I’m in the middle of a 20-week training block targeting four races: HYROX Doubles in Bangkok (March), a half marathon in Singapore (April), then back-to-back HYROX races in Hong Kong (May) - singles on Friday and mixed doubles on Saturday. The entire plan was designed with AI, and I’ve been updating it with real training data as the weeks go by. You can see the full plan at training.vinlam.com.
Why not just hire a coach?
I needed a periodised plan that peaked for four races across three months, with two of them on consecutive days. You can’t just use a cookiecutter plan off the internet . Running volume, strength work, HYROX conditioning, multiple tapers, recovery windows, half marathon sharpening - around a full-time job.
I could have hired a coach. But I already had a clear picture of my goals, training history, and constraints. I just needed help structuring it into a plan.
Writing the metaprompt
Before generating the plan, I used AI to help me write the prompt itself - a metaprompt. I iterated on the structure until it covered everything a coach would need to know.
The final prompt covered: my athlete profile (age, weight, sleep, occupation, training philosophy), performance benchmarks (race results, strength numbers, historical bests), running data (monthly mileage, heart rate zones, pace data), the full race calendar with priorities and target times, my HYROX movement profile (strongest and weakest movements), equipment access across three gyms, training availability and session constraints, recent training logs from the weeks before the plan started, and fixed commitments like travel days.
I also specified the output format: a written strategy explanation plus a structured week-by-week plan.
The back and forth
The first output was decent but not right.
The initial plan had too much running volume in the early weeks and didn’t account for my shin issue. Some gym sessions ran over 50 minutes, which I’d explicitly said I couldn’t do. The HYROX conditioning sessions just said “HYROX circuit” without specifying movements or rep schemes. The taper was too aggressive in one version and too conservative in the next.
Each round I’d point out what was wrong, and it would restructure quickly.
Updating with real data
The plan started in late December 2025. Some weeks I hit it exactly, others I didn’t - an extra rest day here, a session that ran long there. The early weeks are now filled in with what actually happened: real weights, actual paces, specific interval splits.
The more useful part is feeding that data back to the AI and asking it to adjust future weeks. If I hit heavy squats comfortably, it bumps the target up. If my tempo run pace improved, the intervals get faster. The plan progresses based on what I’m actually doing, not what was guessed at the start.
I’ve also been raising weaknesses as they show up. My shoulder endurance is a limiter on wall balls, my quads fade on lunges, and my burpee conditioning needs work. I flag these to the AI and it adds targeted work into upcoming weeks.
Building the dashboard
I built a single-file HTML page to visualise the plan - colour-coded workout types, phase badges, weekly volume stats, with the current week highlighted. Same approach I wrote about in my previous post on building tools with AI: one file, no frameworks, no build steps.
Was it worth it?
A human coach would probably produce a better plan, especially with ongoing feedback. But for a few hours of prompt writing and iteration, I got something more specific to my situation than any template. Writing the prompt also forced me to think through my own priorities, which was useful on its own.
I’m in week 8. Bangkok is four weeks away, Hong Kong is the final test in May. We’ll see how it goes.